| { |
| "File Number": "1017", |
| "Title": "Nichelle and Nancy: The Influence of Demographic Attributes and Tokenization Length on First Name Biases", |
| "Limitation": "Did you describe the limitations of your work? page 5 first section \"Limitations\"\n3 A2. Did you discuss any potential risks of your work? page 5 second section \"Ethics statement\"\n3 A3. Do the abstract and introduction summarize the paper’s main claims? page 1 \"abstract\" and section 1 \"Introduction\"\n7 A4. Have you used AI writing assistants when working on this paper? Left blank. B 3 Did you use or create scientific artifacts? Sections 2, 3, and 4\n3 B1. Did you cite the creators of artifacts you used? Sections 2, 3, and 4\n3 B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix\n3 B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Appendix\nB4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank. B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank. 3 B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix\nC 3 Did you run computational experiments? Sections 2, 3, and 4\n3 C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix\nThe Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance. 3 C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix\n3 C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3\nC4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank. D 7 Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank. D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response. D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants’ demographic (e.g., country of residence)? No response. D3. Did you discuss whether and how consent was obtained from people whose data you’re using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response. D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response. D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response.", |
| "abstractText": "Through the use of first name substitution experiments, prior research has demonstrated the tendency of social commonsense reasoning models to systematically exhibit social biases along the dimensions of race, ethnicity, and gender (An et al., 2023). Demographic attributes of first names, however, are strongly correlated with corpus frequency and tokenization length, which may influence model behavior independent of or in addition to demographic factors. In this paper, we conduct a new series of first name substitution experiments that measures the influence of these factors while controlling for the others. We find that demographic attributes of a name (race, ethnicity, and gender) and name tokenization length are both factors that systematically affect the behavior of social commonsense reasoning models.", |
| "1 Introduction": "Social science studies have shown that individuals may face race or gender discrimination based on demographic attributes inferred from names (Bertrand and Mullainathan, 2004; Conaway and Bethune, 2015; Stelter and Degner, 2018). Similarly, large language models exhibit disparate behaviors towards first names, both on the basis of demographic attributes (Wolfe and Caliskan, 2021) and prominent named entities (Shwartz et al., 2020). Such model behavior may cause representational harms (Wang et al., 2022a) if names associated with socially disadvantaged groups are in turn associated with negative or stereotyped attributes, or allocational harms (Crawford, 2017) if models are deployed in real-world systems, like resume screeners (O’Neil, 2016; Blodgett et al., 2020).\nThe task of social commonsense reasoning (Sap et al., 2019; Forbes et al., 2020), in which models must reason about social norms and basic human psychology to answer questions about interpersonal situations, provides a particularly fruitful setting\nfor studying the phenomenon of name biases in NLP models. Questions in the Social IQa dataset (Sap et al., 2019), for example, describe hypothetical social situations with named, but completely generic and interchangeable, participants (e.g. “Alice and Bob”). Social IQa questions require models to make inferences about these participants, yet they maintain the convenient property that correct (or best) answers should be invariant to name substitutions in most or all cases.\nLeveraging this invariance property, prior work (An et al., 2023) has demonstrated that social commonsense reasoning models acquire unwarranted implicit associations between names and personal attributes based on demographic factors (Fig. 1). Building upon this finding, we investigate a natural follow-up question: why?\nWe identify two possible factors that cause a model’s disparate treatment towards names: demographic attributes and tokenization length. We hypothesize that names associated with different demographic attributes, in particular race, ethnicity, and gender may cause a model to represent and treat them differently. These demographic\n388\nattributes are also strongly correlated with corpus frequency and tokenization length (Wolfe and Caliskan, 2021). Tokenization (or segmentation) breaks down an input sentence into a series of subword tokens from a predefined vocabulary, each of which is then, typically, mapped to a word embedding as the input to a contemporary language model. A name’s tokenization length refers to the number of subwords in the name following tokenization. In this work, we refer to singly tokenized and multiply tokenized names as those consisting of one or multiple tokens after tokenization, respectively. As a result, singly tokenized names are represented with a single embedding vector, while multiply tokenized names are represented by two or more. With these potential confounds, we attempt to address the research question: In social commonsense reasoning, to what extent do demographic attributes of names (race, ethnicity, and gender) and name tokenization length each have an impact on a model’s treatment towards names?\nWe first conduct an empirical analysis to understand the distribution of tokenization lengths in names given demographic attributes, and viceversa. Adopting the open-ended bias-discovery framework, SODAPOP (An et al., 2023), we then analyze the impact of demographic attributes and tokenization length on model behavior. We find that both factors have a significant impact, even when controlling for the other. We conclude that due to correlations between demographics and tokenization length, systems will not behave fairly unless both contributing factors are addressed. Finally, we show that a naïve counterfactual data augmentation approach to mitigating name biases in this task is ineffective (as measured by SODAPOP), concluding that name biases are primarily introduced during pre-training and that more sophisticated mitigation techniques may be required.", |
| "2 Demographic Attributes and Tokenization Length are Correlated": "Previously, Wolfe and Caliskan (2021) have shown that White male names occur most often in pretraining corpora, and consequently, White male names are more likely to be singly tokenized. We replicate this finding by collecting 5,748 first names for 4 races/ethnicities (White, Black, Hispanic, and Asian) and 2 genders (female and male) from a U.S. voter files dataset compiled by Rosenman et al. (2022) (specific data processing and name inclusion criteria are in appendix B.1). We compute and plot the conditional probabilities of tokenization length given demographic attributes (race/ethnicity and gender) and vice-versa in Fig. 2 using the BERT tokenizer (Devlin et al., 2019; Wu et al., 2016). Let ST be the event that a name is singly tokenized. We see in Fig. 2 that P (White|ST ), P (ST |White), P (Male|ST ), and P (ST |Male) are substantially higher than other conditional probabilities involving ST 1, confirming Wolfe and Caliskan (2021).\nThese observations suggest that a model tends to represent White names and male names differently from others in terms of the tokenization length. Given these substantial differences in tokenization lengths across demographic groups, we are motivated to investigate whether tokenization is a primary cause of disparate treatment of names across demographic groups. It is important to note here that, even if tokenization were the primary cause of disparate treatment of names across demographic groups, this discovery would not in itself resolve the fairness concerns of representational and allocational harms based on race, ethnicity and gender, but it might point to possible technical solutions. However, as we will show in the next section, dis-\n1We present similar results for RoBERTa (Liu et al., 2019) and GPT-2 (Radford et al., 2019) tokenizer (Sennrich et al., 2015) in Fig. 6 (appendix A).\nparate treatment of names across demographic attributes persists strongly even when controlling for tokenization length (and vice-versa).", |
| "3 Analyzing the Influences via SODAPOP": "We follow SODAPOP (An et al., 2023) to investigate how the two factors in § 2 influence a Social IQa model’s behavior towards names.", |
| "3.1 Experiment Setup": "SODAPOP leverages samples from Social IQa (Sap et al., 2019), a social commonsense reasoning multiple choice questions (MCQ) dataset. Each MCQ consists of a social context c, a question q, and three answer choices τ1, τ2, τ3, one of which is the only correct answer. An example is shown in Fig. 1.\nSubgroup names For controlled experiments, we obtain at most 30 names for each subgroup categorized by the intersection of race/ethnicity, gender, and tokenization length (BERT tokenizer), resulting in a total of 686 names. Table 1 (appendix) shows the specific breakdown for each group.\nSuccess rate vectors Using millions of MCQ instances, SODAPOP quantifies the associations between names and words using success rate vectors (SR vectors): a vector whose entries are the probability of a distractor τi containing word w to fool the model, given that name n is in the context. For illustration, out of 5,457 distractors containing the word “violent” we generated for the name “Nichelle” (Fig. 1), 183 misled the model to pick the distractor over the correct answer choice. The success rate for the word-name pair (“violent”, “Nichelle”) is 1835457 = 3.28%. We present more details, including the formal mathematical definition of success rate, in appendix B.2.\nClustering of the success rate vectors The clustering of SR vectors can be visualized by tSNE projections. To quantify the tightness of clustering\nbetween two groups of SR vectors A,B, we first find the centroids −→cA,−→cB by averaging 3 random SR vectors within each group. Then, for each SR vector −→s (including the 3 random vectors for centroid computation), we assign a label a if its euclidean distance is closer to −→cA, otherwise b. We check the accuracy x of this naïve membership prediction. The membership prediction accuracy on SR vectors produced by a fair model would be close to 0.5, indicating that name attributes are not easily recoverable from their corresponding SR vectors. We evaluate the statistical significance using a variant of the permutation test. The null hypothesis is that the SR vectors of groups A and B are no more clusterable than a random re-partitioning of A ∪B would be. We randomly permute and partition the SR vectors into A′, B′ with the same cardinality each and relabel them. We predict the membership of SR vectors based on their distance to the new centroids −→cA′,−→cB ′, obtaining accuracy x′. The p-value P (x′ > x) is estimated over 10, 000 runs.", |
| "3.2 Results: Both Factors Matter": "We use the 686 names across all subgroups, almost evenly distributed by demographic attributes, and obtain the tSNE projection of their SR vectors (obtained using BERT, and the dimension is 736) in Fig 3. We observe clear clustering by tokenization length, race/ethnicity, and gender. Since tokenization length is generally correlated with corpus frequency, we also see weak clustering of the SR vectors by frequency.\nWe report the membership prediction accuracy of SR vectors (obtained by running SODAPOP on a finetuned BERT model for Social IQa) for all pairs of subgroups in Fig. 4a. Each cell in the figure shows the separability of SR vectors for names from two groupings. To illustrate, the top left cell shows singly tokenized White male names are highly separable (> 80%) from singly tokenized White female names; the entire heatmap shows the\nresults for all pairs. As we vary one and control the other confounding factors, we find that each of race/ethnicity, gender, and tokenization length are name attributes that lead to systematically different model behavior, as measured by membership prediction accuracy. Almost all prediction accuracy is close to 1.0, indicating perfect separation of the clusters, with p < 0.001 in nearly all settings. We see in Fig. 4a, for instance, that SR vectors of singly tokenized Black female names and singly tokenized White female names are perfectly classified, so race is still a pertinent factor even controlling for gender and tokenization. In contrast, SR vectors for singly tokenized Asian male and Asian female names are not distinguishable, although gender appears to influence model behavior under most other controlled settings.\nWe obtain experimental results for RoBERTa and GPT-2 in appendix C. We observe that these additional results also demonstrate a similar trend as BERT, generally supporting the hypothesis that models exhibit disparate behavior for different names based on their demographic attributes as well as tokenization length. However, the results for RoBERTa and GPT-2 are less strong than that of BERT. We speculate a variety of reasons that could give rise to the different results among these models. One potential major cause is the different tokenization algorithms used by the models: BERT uses WordPiece (Wu et al., 2016) while RoBERTa and GPT-2 use Byte-Pair Encoding (Sennrich et al.,\n2015) for tokenization. Due to this difference, the tokenization length of a name can vary in these models. For example, “Nancy” is singly tokenized in BERT but is broken down into [“N”, “ancy”] in RoBERTa or GPT-2. Beyond tokenization, the different pre-training algorithms and training corpora will also likely contribute to the slightly different observations between Fig. 4 and Fig. 10.", |
| "4 Counter-factual Data Augmentation": "We apply counter-factual data augmentation (CDA) to the Social IQa training set as we attempt to finetune a model that is indifferent to both tokenization length and the demographic attributes of names. We choose to experiment with CDA because it would shed light on the source of name biases. If biases mostly arise from finetuning, we expect finetuning on Social IQa with CDA would largely address the problem; otherwise, biases mostly originate from pre-training and are not easily overridden during finetuning.\nFor each Social IQa sample, we identify the original names using Stanford NER (Finkel et al., 2005). We find that more than 99% of samples contain one or two names. We create copies of the MCQ samples and replace the identified names with random names from our sampled sub-groups such that the overall name frequency is evenly distributed over tokenization lengths and demographic attributes, resulting in an augmented set whose size increases by 16×. We finetune a BERT model\nwith the augmented set (details in appendix B.2). However, this naïve solution is rather ineffective (Fig. 4b). This negative result is not surprising as it aligns with the observations that SODAPOP could detect biases even in models debiased with state-of-the-art algorithms (An et al., 2023). It also indicates that pre-training contributes to the biased model behavior. Hence, a more sophisticated solution is needed to tackle this problem.", |
| "5 Related Work": "Social biases in language models Multiple recent works aim to detect social biases in language models (Rudinger et al., 2018; Zhao et al., 2018, 2019; Nangia et al., 2020; Li et al., 2020; Nadeem et al., 2021; Sap et al., 2020; Parrish et al., 2022). Some works specifically diagnose biases in social commonsense reasoning (Sotnikova et al., 2021; An et al., 2023), but they do not explain what causes a model to treat different names dissimilarly; in particular, these works do not consider the influence of tokenization length on model behavior towards different names.\nName artifacts Previous research indicates that language models exhibit disparate treatments towards names, partially due to their tokenization or demographic attributes (Maudslay et al., 2019; Czarnowska et al., 2021; Wang et al., 2022b). However, thorough analyses of the factors influencing first name biases are lacking in these works. While Wolfe and Caliskan (2021) study the systematic different internal representations of name embeddings in language models due to the two factors, we systematically study how the two factors both connect with the disparate treatment of names by a model in a downstream task.", |
| "6 Conclusion": "We have demonstrated that demographic attributes and tokenization length are both factors of first names that influence social commonsense model behavior. Each of the two factors has some independent influence on model behavior because when controlling one and varying the other, we observe disparate treatment of names. When controlling for tokenization length (e.g. Black male singlytokenized names vs White male singly-tokenized names) we still find disparate treatment. Conversely, when we control for demographics (e.g. Black female singly-tokenized vs Black female\ntriply-tokenized names), the model also treats those names differently. Because demographic attributes (race, ethnicity, and gender) are correlated with tokenization length, we conclude that systems will continue to behave unfairly towards socially disadvantaged groups unless both contributing factors are addressed. We demonstrate the bias mitigation is challenging in this setting, with the simple method of counterfactual data augmentation unable to undo name biases acquired during pre-training.\nLimitations\nIncomplete representation of all demographic groups We highlight that the names used in our study are not close to a complete representation of every demographic group in the United States or world. In our study, we adopt the definition of race/ethnicity from the US census survey, using US-centric racial and ethnic categorizations that may be less applicable in other countries. We adopt a binary model of gender (female and male), based on the SSA dataset, which is derived from statistics on baby names and assigned sex at birth; this approach limits our ability to study chosen first names, or to study fairness with respect to nonbinary and transgender people. For race/ethnicity, our study is limited to US census categories of White, Black, Hispanic, and Asian. We are unable to include American Indian or Alaska Native in our study, for instance, as we were unable to identify any names from this group that met our inclusion criteria of > 50% membership according to our name data source.\nFurthermore, by using first names as a proxy for demographic attributes, we are only able to study certain demographic attributes that plausibly correlate with names (e.g., race, ethnicity, and gender) but not other demographic attributes that are likely harder to infer from names (e.g., ability or sexual orientation). Other demographic attributes that may be discernible to varying degrees from first names were excluded from the scope of this study (e.g., nationality, religion, age).\nAssumption: Invariance under name substitution Invariance under name substitution, while a valuable fairness criterion for Social IQa, may not hold in all other task settings. For example, a factoid QA system should provide different answers to the questions “What year was Adam Smith born?” (1723) and “What year was Bessie Smith born?” (1894).\nExtended evaluation time and heavy computational costs Due to the huge number of MCQ instances we construct for evaluation and a diverse set of names to cover multiple demographic identities, it takes a considerably large amount of time and computational resources to obtain the analysis results. We detail the approximated time and computational budget in appendix B.2. However, it is worth noting that the extensive analysis on a wide range of MCQ instances and names makes our observations more statistically robust. A future research direction may be optimizing the implementation of SODAPOP framework, which we use as a major experiment setup to obtain the analysis, for more efficient evaluation.\n(In)effectiveness of counter-factual data augmentation It is worth noting that the ineffective result we obtained is not surprising because SODAPOP has demonstrated that models that are trained with existing state-of-the-art debiasing algorithms continue to treat names differently (An et al., 2023). Although we find that controlling the name distribution in the finetuning dataset to be rather ineffective in mitigating the disparate treatment of names, it is an open question if applying CDA to the pre-training corpus would be more effective. A recent work proposes to apply CDA to the pre-training corpus (Qian et al., 2022), and it will likely be a great source to use for investigating our open question here.\nEthics Statement\nPotential risks Our paper contains an explicit example of demographic biases in a social commonsense reasoning model (Fig. 1). This observation does not reflect the views of the authors. The biased content is for illustration purpose only. It should not be exploited for activities that may cause physical, mental, or any form of harm to people.\nThe potential benefits from our work include: (1) insights into the factors that influence a social commonsense reasoning model’s behavior towards first names; (2) the potential for increased awareness of these factors to encourage more cautious deployment of real-world systems; and (3) better insights into the challenges of debiasing, and how demographic and tokenization issues will both need to be addressed.\nDifferences in self-identifications We have categorized names into subgroups of race/ethnicity\nand gender by consulting real-world data as we observe a strong statistical association between names and demographic attributes (race/ethnicity and gender). However, it is crucial to realize that a person with a particular name may identify themselves differently from the majority, and we should respect their individual preferences and embrace the differences. In spite of the diverse possibilities in selfidentification, our observations are still valuable because we have designed robust data inclusion criteria (detailed in appendix B.1) to ensure the statistical significance of our results.", |
| "Acknowledgements": "We thank the anonymous reviewers for their constructive feedback. We also thank Neha Srikanth, Abhilasha Sancheti, and Shramay Palta for their helpful suggestions to improve the manuscript.", |
| "A Additional Analysis on Frequency, Tokenization, and Demographic Attributes of Names": "We provide the complementary plots for Fig. 2 by showing the raw counts of the names in Fig. 5. We also present preliminary observations on the connection between frequency, tokenization, and demographic attributes of names for RoBERTa and GPT-2 tokenizer in this section. Theses results (Fig. 6) are similar to those in § 2. White male names are more likely to be singly tokenized in RoBERTa or GPT-2 as well. We observe that the conditional probability that a name is singly tokenized given that it is Asian is also quite high. We speculate the reason for this is that Asian names have fewer characters in their first names on average (4.40) compared to that of Black names (6.48) and Hispanic names (6.41), which cause Asian names to be more likely singly tokenized as well.\nIn addition, we count the occurrence of 608 names (a subset of the 5,748 names in § 2) in Wikipedia2 and BooksCorpus (Zhu et al., 2015), which are used to pre-train BERT and RoBERTa. Fig. 7 illustrates the distribution of name frequency over different tokenization lengths. We see that, regardless of the model, most singly tokenized names have higher average frequency, whereas multiply tokenized names share similar distributions with lower frequency overall.", |
| "B Detailed Experiment Setup": "B.1 Experiments for Preliminary Observations\nNames We collect people’s first names from a U.S. voter files dataset compiled by Rosenman et al. (2022). We filter out names whose frequency in the dataset is less than 200. Since each name is not strictly associated with a single race/ethnicity, but rather reflects a distribution over races/ethnicities, we analyze only names for which the percentage of people with that name identifying as that race/ethnicity is above 50%. We assign a binary gender label to each name by cross-referencing gender statistics in the SSA dataset.3 If the name is absent from the SSA dataset, we omit that name.\n2https://huggingface.co/datasets/wikipedia 3https://www.ssa.gov/oact/babynames/\nWith these constraints, there is only one name for the category “Other race/ethnicity\". For robust statistical analysis, we choose not to include this category but only the other four categories in the data source, which are White, Black, Hispanic, and Asian. There is a total of 5,748 names.\nModels We use three popular language models for the analysis, namely BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and GPT-2 (Radford et al., 2019). BERT uses WordPiece (Wu et al., 2016) for tokenization, while both RoBERTa and GPT-2 use Byte-Pair Encoding (Sennrich et al., 2015) as their tokenization algorithm. BERT-base has 110 million parameters. RoBERTa-base has 123 million parameters. GPT-2 has 1.5 billion parameters. No finetuning is needed for experiments in § 2 because tokenization of input is invariant to\nfurther finetuning in a downstream task.\nB.2 Experiments with SODAPOP\nSocial IQa To examine machine intelligence in everyday situations, Sap et al. (2019) publish a social commonsense reasoning multiple-choice\ndataset Social IQa. Each MCQ consists of a social context, a question, and three answer choices, one of which is the only correct answer. An example from Social IQa is Context: “Kai made a wish and truly believed that it would come true.” Q: “How would you describe Kai?” A1: “a cynical person” A2: “like a wishful person” A3: “a believing person” (correct choice). There are 33, 410 samples in the training set and 1, 954 instances in the development set.\nGenerating distractors To detect a model’s disparate treatment towards names, SODAPOP substitutes the name in a MCQ sample with names associated with different races/ethnicities and genders, and generate a huge number of new distractors to robustly test what makes a distractor more likely to fool the MCQ model, thus finding the model’s implicit associations between names and attributes. We follow the same algorithm proposed by An et al. (2023) to generate distractors using a masked-token prediction model (RoBERTa-base). We generate distractors from the correct choice of 50 MCQ samples in Social IQa (Sap et al., 2019). We utilize the same list of names for distractor generation as in SODAPOP. In our study, we take the union of all the distractors generated with different names for a context to form new MCQ samples for more robust\nresults. The total number of MCQ constructed via this step is 4,840,776.\nSuccess rate Recall that each MCQ in Social IQa consists of a social context c, a question q, and three answer choices τ1, τ2, τ3, one of which is the only correct answer. Formally, for an arbitrary distractor τi, the success rate of a word-name pair (w, n) is\nSR(w, n) = P ( argmax j∈{1,2,3} M(c, q, τj) = i ∣∣∣∣\n( w ∈ tok(τi) ) ∧ ( n ∈ tok(c) )) , (1)\nwhere M(c, q, τj) produces the logit for answer choice τj using a MCQ model M, and tok splits the input by space so as to tokenize it into a bag of words and punctuation. A success rate vector for a name n composes |V | entries of SR(w, n) for all w ∈ V , where V is the set of vocabulary (i.e., words appearing in all distractors above a certain threshold). Specifically, we set the threshold to be 1, 000 in our experiments.\nModels We conduct experiments using three popular language models, namely BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and GPT2 (Radford et al., 2019). The size of each model is specified in appendix B.1. We finetune each model on the Social IQa training set with a grid search for hyperparameters (batch size = {3, 4, 8}, learning rate = {1e−5, 2e−5, 3e−5}, epoch = {2, 4, 10}). Although different hyper-parameters lead to varying final performance on the development set of Social IQa, we find them to be within a small range in most cases (within 1%− 2%). Since our analysis does not highly depend on the performance of a model, we arbitrarily analyze a model that has a decent validation accuracy among all. In our study, the BERT-base model is finetuned with batch size 3, learning rate 2e−5 for 2 epochs and achieves 60.51% on the original dev set. The RoBERTabase model is finetuned with batch size 8, learning rate 1e−5 for 4 epochs and achieves 70.51% on the original dev set. The GPT-2 model is finetuned with batch size 4, learning rate 2e−5 for 4 epochs and achieves 61.91% on the original dev set. To finetune on the counter-factually augmented dataset, we conduct grid search for batch size = {2, 3, 8}, learning rate = {1e−5, 2e−5} for 1 epoch. We obtain similar dev set accuracy for these setting, all about 60%.\nThe evaluation time for 4 million MCQs across more than 600 names is costly. We approximate that it takes about 7 days using 30 GPUs (a combination of NVIDIA RTX A4000 and NVIDIA TITAN X) for each model. However, we note that a smaller number of MCQ instances and names may sufficiently capture the biased behavior of a model. We choose to include an extremely large number of test instances and a wide range of names to ensure the robustness of our study. Although important, it is out of the scope of this paper to find the optimal size of the bias-discovery test set to minimize computation time and resources.\nSubgroup names For fine-grained analysis that compares a model’s different behavior towards two name groups that only vary by one confounding factor, we compile subgroups of names that share the same race/ethnicity, gender, and tokenization length. For example, White female names with tokenization length 2 is one subgroup of names. In total, we sample 686 names for BERT and 608 names for RoBERTa and GPT-2. Table. 1 shows the specific number of names in each subgroup. Given the data source available to us, we are unable to collect an enough number of names for certain subgroups (e.g., Asian male names with tokenization length 3). Nonetheless, these limitations do not affect our findings of the different treatment towards other subgroups with a sufficiently large number of names.", |
| "C Additional Experiment Results": "We illustrate the tSNE projections of SR vectors for RoBERTa and GPT-2 in Fig. 8 and Fig. 9 respectively. The dimension of the SR vectors is 660 for these two models. The plots show that, as we control each of the factors in our analysis, both RoBERTa and GPT-2 treat names differently in the downstream task of social commonsense reasoning.\nWe also report the membership prediction accuracy for RoBERTa and GPT-2 in Fig. 10. We observe that gender, race/ethnicity, and tokenization length are all strongly correlated with the model’s disparate treatment of names in these models as well. GPT-2 behaves similarly as BERT, where tokenization length, race/ethnicity, and gender are all factors that indicate the model’s different behavior towards names.", |
| "D Responsible NLP": "Licenses We have used BERT, RoBERTa, and GPT-2 for our empirical studies. BERT uses Apache License Version 2.0,4 and both RoBERTa and GPT-2 use MIT License.5 We are granted permission to use and modify these models for our experiments per these licenses.\nWe also use Stanford NER in our experiments, which is under GNU General Public License (V2 or later).6\nThe pipeline SODAPOP is under AttributionNonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).7 We have the permission to copy and redistribute the material in any medium or format.\nThe dataset Social IQa is under Creative Commons Attribution 4.0 International License8 as it was published by Association for Computational Linguistics. Per the license, we may “copy and redistribute the material in any medium or format” and “remix, transform, and build upon the material for any purpose, even commercially.”\nThe first name dataset (Rosenman et al., 2022) is under CC0 1.0 Universal (CC0 1.0) Public Domain Dedication.9 Everyone can copy, modify, distribute and perform the work, even for commercial purposes, all without asking permission.\nConsistency with the intended use of all artifacts We declare that the use of all models, datasets, or scientific artifacts in this paper aligns with their intended use.\n4https://www.apache.org/licenses/LICENSE-2.0 5https://opensource.org/licenses/MIT 6https://www.gnu.org/licenses/old-licenses/\ngpl-2.0.html 7https://creativecommons.org/licenses/ by-nc-nd/4.0/ 8https://creativecommons.org/licenses/by/4.0/ 9https://creativecommons.org/publicdomain/ zero/1.0/\nACL 2023 Responsible NLP Checklist", |
| "A For every submission:": "3 A1. Did you describe the limitations of your work?\npage 5 first section \"Limitations\"\n3 A2. Did you discuss any potential risks of your work? page 5 second section \"Ethics statement\"\n3 A3. Do the abstract and introduction summarize the paper’s main claims? page 1 \"abstract\" and section 1 \"Introduction\"\n7 A4. Have you used AI writing assistants when working on this paper? Left blank.\nB 3 Did you use or create scientific artifacts? Sections 2, 3, and 4\n3 B1. Did you cite the creators of artifacts you used? Sections 2, 3, and 4\n3 B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Appendix\n3 B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Appendix\nB4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Not applicable. Left blank.\nB5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Not applicable. Left blank.\n3 B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Appendix\nC 3 Did you run computational experiments? Sections 2, 3, and 4\n3 C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Appendix\nThe Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.\n3 C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Appendix\n3 C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Section 3\nC4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Not applicable. Left blank.\nD 7 Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.\nD1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.\nD2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants’ demographic (e.g., country of residence)? No response.\nD3. Did you discuss whether and how consent was obtained from people whose data you’re using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.\nD4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.\nD5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response." |
| } |
